UC Berkeley: Random Forests and vLLM Research

A California research university whose AI contributions include tree-ensemble methods and the Berkeley-developed vLLM system for managing language-model inference memory.

kindUniversity
founded1868

Updated

Statistical learning and model-serving systems

Berkeley dates the University of California's foundation to 1868. Its AI record spans statistical methods and the systems needed to use trained models, rather than only the development of chat applications.

Leo Breiman and Adele Cutler's university-hosted random forests documentation describes ensembles of trees whose votes produce classifications, with methods for estimating error and variable importance. The combination provides a method for classification and for inspecting which input variables affect predictions.

The June 2023 vLLM release explicitly identifies UC Berkeley as its development site and reports prior deployment in Chatbot Arena and the Vicuna Demo. PagedAttention manages attention key-value caches in blocks, reducing memory-management waste during inference. The release's throughput ratios are results from its specified experiments, rather than guarantees for every workload.

Sources

  1. [1]
    About Berkeley

    University of California Berkeley · retrieved Oct 24, 2026

  2. [2]
    Random forests - classification description

    Leo Breiman and Adele Cutler, UC Berkeley · retrieved Oct 24, 2026

  3. [3]

University of California Berkeley history

2023

│
vLLM publicly releases its language-model serving engine

With vLLM

The UC Berkeley project released an open-source inference engine using PagedAttention to reduce wasted attention-cache memory and batch more requests. Its authors reported throughput gains on LLaMA models; these were their own tests, not a universal speedup.

vLLM
Open source

2001